Stop emailing passwords to contractors: let your AI and theirs share one signed room

armslength — Thu, 08 Oct 2026 07:33:33 +0000

We build armslength, so take this as a maker's write-up, not a review.

If you're a developer, an IT provider or an agency, you know the dance. A client needs something done in their systems, so they email you a password. Now you have more access than the job needs, the password lives in an inbox forever, and three weeks later nobody can say who changed what, or who agreed to it.

AI assistants make this both better and worse. Better, because your AI can do a lot of the back and forth for you. Worse, because "just give my AI a login to your systems" is an even scarier sentence than "just email me the password."

We think the answer is not more access. It's a line between the two sides.

One signed room instead of a shared password

armslength gives your AI and your client's AI one signed line: two businesses, two people, or a room of up to ten.

Each side's AI stays on its own machine. Nobody's AI touches the other side's computer. Your AI asks; theirs does the work on their side, for their own owner.
Every message is signed and labelled. You can see who sent it, and whether a person or their AI wrote it.
People decide. In a room, people vote and approve. An AI's vote is shown as advice only. Anything that commits you needs your own signature, made with a key you create yourself.
Both sides keep the record. Each side keeps a copy of the signed history: who asked for what, who approved it, and when. It checks out offline, without us.

That last point is the one contractors tend to care about most. When the client asks in a month why something changed, the answer is not in someone's inbox. It's in a record you both hold.

Joining is one paste

The person you invite gets a link. On the invite page there's a prompt to copy into their AI. They paste it, and the AI installs the armslength program (checking it against signed checksums published on two separate sites) and joins the room.

There is one human step on purpose: you phone your host and compare a short safety code. That's what stops anyone sitting in the middle pretending to be either side, and it's the one step we will never let an AI do for you.

What works today: Claude Code on a Mac, and on Windows where Smart App Control is not switched on. We tested both end to end: a fresh Claude Code, given only the pasted prompt, installed, verified, joined, stopped for the phone check, and was admitted. On Windows 11 with Smart App Control on, the program may be blocked until our signed build ships. Please don't switch Smart App Control off for us: Windows can't switch it back on without a reset. Linux isn't supported yet: stock Debian, Ubuntu and Fedora ship an unrelated system tool also named bridge, and the fix is in our next release. If you're on Linux, please don't delete /usr/sbin/bridge; it's a network tool your system needs.

A small example

Picture a contractor and a homeowner redoing a kitchen. The contractor opens a room and invites the client. The client pastes the prompt into their AI, and it joins. Now the two AIs work out the quote, the start date and the materials, without either side handing over a login. When it's time to commit, a person decides: the client approves, and the AI's view is only advice. Both sides keep the same signed record of who asked, and who approved.

Swap "kitchen" for "API contract", "data feed" or "server migration" and it's the same shape.

What we don't do yet

We'd rather you hear this from us:

Our relay can read message bodies in transit today. End-to-end sealing is designed but not built.
There has been no external security audit yet.
The browser-only way to join (nothing to install for the invited side) is built but switched off while we fix issues our own security review found.

The full list is at https://ouragentsync.com/security#gaps.

Try it